ETCHR: Editing To Clarify and Harness Reasoning

Hugging Face Daily Papers Papers
image-editing visual-reasoning multimodal-llm chain-of-thought reasoning-enhancement qwen3 gemini

Summary

ETCHR is a novel image editing approach that decouples visual reasoning from image generation, using a two-stage training process (Reasoning Imitation and Reasoning Enhancement) to improve multimodal language model performance across five visual reasoning tasks. It achieves consistent gains of 4-5% Pass@1 on models like Qwen3-VL-8B, Gemini-3.1-Flash-Lite, and Kimi K2.5.

Multimodal Large Language Models have advanced visual reasoning, yet a purely textual chain of thought remains a bottleneck for questions that require fine-grained focus or view transformations. The ''think with images'' paradigm narrows this gap, but existing approaches are either constrained by fixed predefined toolkits or produce noisy intermediate images from unified multimodal methods. We pursue a third option: using a dedicated image editing model and decouple it with an understanding model. However, off-the-shelf image editors fail as reasoning assistants with two complementary gaps: a language-side gap, where editors trained as passive instruction-followers cannot map an abstract question to an appropriate visual transformation, and a generation-side gap, where edit correctness degrades as reasoning depth grows. Guided by this analysis, we introduce ETCHR (Editing To Clarify and Harness Reasoning), a question-conditioned, reasoning-aware image editor decoupled from the downstream understanding model and trained with a two-stage recipe targeted at the two gaps: Reasoning Imitation via supervised fine-tuning on edit trajectories, followed by Reasoning Enhancement with VLM-derived rewards for edit correctness and downstream reasoning accuracy. Since the editor is decoupled, ETCHR plugs into different open- and closed-source MLLMs in a training-free manner. Across five task families (fine-grained perception, chart understanding, logic reasoning, jigsaw restoration, and 3D understanding), ETCHR raises average Pass@1 from 55.95 to 60.77 (+4.82) with Qwen3-VL-8B, from 65.08 to 70.55 (+5.47) with Gemini-3.1-Flash-Lite, and from 76.55 to 81.16 (+4.61) with the 1T-parameter MoE model Kimi K2.5.
Original Article
View Cached Full Text

Cached at: 05/25/26, 02:35 AM

Paper page - ETCHR: Editing To Clarify and Harness Reasoning

Source: https://huggingface.co/papers/2605.23897

Abstract

A novel image editing approach called ETCHR is introduced that decouples visual reasoning from image generation, improving multimodal language model performance across multiple visual reasoning tasks through a two-stage training process.

Multimodal Large Language Modelshave advancedvisual reasoning, yet a purely textualchain of thoughtremains a bottleneck for questions that require fine-grained focus or view transformations. The ‘’think with images’’ paradigm narrows this gap, but existing approaches are either constrained by fixed predefined toolkits or produce noisy intermediate images from unified multimodal methods. We pursue a third option: using a dedicatedimage editing modeland decouple it with an understanding model. However, off-the-shelf image editors fail as reasoning assistants with two complementary gaps: a language-side gap, where editors trained as passive instruction-followers cannot map an abstract question to an appropriate visual transformation, and a generation-side gap, where edit correctness degrades as reasoning depth grows. Guided by this analysis, we introduce ETCHR (Editing To Clarify and Harness Reasoning), a question-conditioned,reasoning-aware image editordecoupled from the downstream understanding model and trained with a two-stage recipe targeted at the two gaps:Reasoning Imitationvia supervised fine-tuning on edit trajectories, followed byReasoning EnhancementwithVLM-derived rewardsfor edit correctness and downstream reasoning accuracy. Since the editor is decoupled, ETCHR plugs into different open- and closed-source MLLMs in a training-free manner. Across five task families (fine-grained perception, chart understanding, logic reasoning, jigsaw restoration, and 3D understanding), ETCHR raises averagePass@1from 55.95 to 60.77 (+4.82) withQwen3-VL-8B, from 65.08 to 70.55 (+5.47) withGemini-3.1-Flash-Lite, and from 76.55 to 81.16 (+4.61) with the 1T-parameter MoE modelKimi K2.5.

View arXiv pageView PDFGitHub6Add to collection

Get this paper in your agent:

hf papers read 2605\.23897

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2605.23897 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2605.23897 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2605.23897 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Is This Edit Correct? A Multi-Dimensional Benchmark for Reasoning-Aware Image Editing

Hugging Face Daily Papers

This paper introduces RE-Edit, a benchmark for evaluating image editing systems across five reasoning dimensions (physical, environmental, cultural, causal, referential) to assess logical consistency beyond visual plausibility. The benchmark includes 1,000 samples and evaluates ten open-source and two commercial models, showing that even advanced systems struggle with implicit multi-dimensional reasoning.

STRIDE-ED: A Strategy-Grounded Stepwise Reasoning Framework for Empathetic Dialogue Systems

arXiv cs.CL

STRIDE-ED is a strategy-grounded reasoning framework for empathetic dialogue systems that uses structured multi-stage reasoning combined with a data refinement pipeline and two-stage training (supervised fine-tuning + multi-objective RL) to improve emotional understanding and response generation. The framework demonstrates consistent improvements across open-source LLMs on both automatic metrics and human evaluations.